2026.05.04 AI News Briefing — Upstage raises 560 billion won, industry ministry launches AI Factory program: 'sovereign AI' ignites

Over the long holiday weekend, the 'sovereign AI' front in Korea's AI industry came into sharper focus. Upstage raised 560 billion won, including the first direct investment from the National Growth Fund, and the industry ministry opened a new 52.75 billion won call for its 'AI Factory pilot program'. Globally, SoftBank's plan to spin off a $100 billion AI and robotics company, 'Roze', came to light, and Zhipu's Kimi K2.6 beat GPT-5.5 and Claude on coding benchmarks.

---

1. Upstage raises 560 billion won, including the National Growth Fund's 'first direct AI model investment'

• Source: AI Times / Upstage | 2026.04.30–05.02
• What happened: Korean LLM startup Upstage was selected as the National Growth Fund's 'first direct investment in the AI model sector', raising a total of 560 billion won including 100 billion won of direct government investment. In the same period, the government unveiled plans to put semiconductor infrastructure on the scale of 15,000 GPUs into a 'National AI Computing Center', dubbing it the 'AI highway'. The Financial Services Commission also announced a second mega-project in biotech, following the first one centered on semiconductors and AI.
• Why it matters: The government has entered industrial-policy mode, directly cultivating a single champion for a 'sovereign AI model'. The center of gravity in Korea's LLM ecosystem looks set to shift quickly from 'many competing startups' to 'a lead champion plus public computing infrastructure'.

2. Industry ministry opens a new 52.75 billion won 'AI Factory pilot program' call — targeting chips, autos, and shipbuilding

• Source: Ministry of Trade, Industry and Resources / AI Times | 2026.04–05
• What happened: The Ministry of Trade, Industry and Resources opened a new 52.75 billion won call for its 'AI Factory pilot program', which brings AI into the manufacturing floors of core industries such as semiconductors, automobiles, and shipbuilding. It follows the 90 billion won 'Smart Green Industrial Complex (M.AX·GX) combined call' that began on April 29. Productivity gains, lower defect rates, and shorter lead times are specified as the key evaluation metrics. In the same period, LG CNS reported Q1 revenue of 1.31 trillion won and operating profit of 94.2 billion won (+8.6%/+19.4% year on year), with double-digit growth in its AI and cloud businesses.
• Why it matters: The idea that 'AI makes money in the factory, not the chatbot' was confirmed simultaneously in government funding and in domestic IT services results. Expect a concentration of related solution and contract announcements at 'AI EXPO KOREA 2026' at COEX on May 6–8.

3. SoftBank pursues a spin-off of AI and robotics company 'Roze' — eyeing an IPO of up to $100 billion

• Source: Financial Times / Fortune | 2026.04.30
• What happened: SoftBank is reported to be planning to spin off a new subsidiary, 'Roze', combining AI and robotics, and list it in the US. The target valuation is up to $100 billion, making it a candidate for the largest AI IPO ever. Roze focuses on the 'physical build-out' of AI infrastructure, bundling robots for data center construction with SoftBank's energy, land, and digital infrastructure assets. The IPO could come in the second half of 2026 and is seen as a way to manage SoftBank's enormous financial commitments to OpenAI and large infrastructure projects.
• Why it matters: 'Physical AI' is being carved out as a standalone category in the IPO market. It is likely to change the valuation benchmarks for the robotics and data center investment strategies of Korean conglomerates such as SK, Hyundai, and LG.

4. China's Zhipu 'Kimi K2.6' beats Claude, GPT-5.5, and Gemini on coding benchmarks

• Source: TLDL AI News | 2026.05
• What happened: Results were published showing that Chinese lab Zhipu AI's next-generation model 'Kimi K2.6' outperformed Anthropic's Claude, OpenAI's GPT-5.5, and Google's Gemini on programming challenges. In the same period, IBM announced that its 8-billion-parameter 'Granite 4.1' matches the performance of its previous 32-billion-parameter MoE model, and OpenAI published a postmortem attributing GPT-5.1's tendency to overproduce words like 'goblin' and 'gremlin' to the reward signal used when training a 'nerdy' persona. Security researchers also reported finding 'Shai-Hulud'-themed malware in the PyTorch Lightning training library.
• Why it matters: We have entered an era where the top of the coding and general benchmarks effectively changes every quarter. For Korean companies' LLM strategy, designing a 'model portfolio plus usage policy' now matters more than 'single-model lock-in'.

5. Mercor data breach — 4 TB of voice samples from 40,000 AI contractors leaked

• Source: TLDL AI News | 2026.04.28
• What happened: AI data labeling and contractor-matching platform Mercor suffered a major breach in which roughly 4 TB of voice samples were leaked, affecting about 40,000 contractors. Voice data is a core asset for training next-generation speech and multimodal AI, putting the security risks of the AI training data ecosystem back under scrutiny.
• Why it matters: AI training data is becoming a new target for 'supply chain attacks'. Korean companies should now build security due diligence on data processors into their standard procedures for outsourced labeling, synthetic voice training, and the like.

6. Anthropic releases cybersecurity-focused 'Mythos' model on a limited basis — White House reviews federal reinstatement

• Source: Fox News / Forbes / LLM Stats | 2026.04–05
• What happened: Anthropic released 'Claude Mythos', a new model specialized in cybersecurity capabilities, on a limited basis, judging that some advanced cyber capabilities are too risky for broad market release. Meanwhile, the White House is drafting guidance that would let federal agencies work with Anthropic again, reportedly including a review of access to Mythos. Separately, Anthropic's annualized revenue (ARR) is tallied at about $19 billion, nearly doubling in two months.
• Why it matters: The trend of per-model 'use policies' translating directly into market access is hardening. In Korea's public, defense, and financial sectors, reviewing 'model cards plus use policies' is likely to become a de facto mandatory procurement step.

---

Today's insight

• 'Sovereign AI' has become the top industrial policy priority: It is no coincidence that Upstage's 560 billion won raise, 15,000 GPUs for the National AI Computing Center, and the industry ministry's 52.75 billion won AI Factory program were announced in the same week. The government has started running three axes at once: a champion model, public computing, and factory-floor deployment. Over the next year, deals, acquisitions, and talent moves in Korea's AI market are likely to explode around these three axes.
• The model race is speeding up to a 'quarterly' cadence: Kimi K2.6, IBM Granite 4.1, and OpenAI's GPT-5.1 postmortem — top-tier, small-model, and debugging developments all broke within days of each other. Single-vendor lock-in is now an obvious business risk. Korean companies' LLM strategy should shift its center of gravity to 'model portfolio plus context and cache cost optimization'.
• AI security and governance risk expands to the 'supply chain': The Mercor breach, the PyTorch Lightning malware, and the limited release of Anthropic's Mythos are different scenes of the same story. Governance now covers not just the safety of the model itself but the security of the entire chain: training data, libraries, and outsourced contractors.